Skip to content
hw.dev
hw.dev/signal/smart-test-data-chain-traceability-teradyne
SignalSemiconductor Engineering

Smart Test Has an Algorithm. What It Lacks Is a Data Chain.

Smart test at AI-accelerator scale is not blocked by ML model quality. It is blocked by data latency, missing metadata, and traceability gaps across fab, test, packaging, and field that prevent any model from acting on the right device at the right insertion with enough context to be trusted.

#testing#verification#tools#ai-hardware
Read Original

The roadmap for smart test has been pointing toward adaptive flows, ML-driven binning, and shorter test times for years. Semiconductor Engineering's recent deep dive finds the roadmap is hitting a different wall than the one the vendors are selling against: the ML models exist, but the data infrastructure connecting fab metrology, wafer sort, package assembly, burn-in, final test, system-level test, and field monitoring does not. A device can pass every checkpoint and still carry a latent defect that the test record never captured, because the record has missing metadata, inconsistent device identity, or model outputs that cannot be traced back to a physical root cause.

The mechanism is latency. Teradyne's Eli Roth puts it plainly: test cells have historically been optimized to cache data for throughput, not to provide actionable data in seconds. AI-scale devices with 1,200-amp power requirements during test, massively replicated cores, and multi-chiplet interconnects cannot wait hours for a model to act on process history from the fab. The constraint being removed in the next investment cycle is not model quality but data plumbing: getting the right process context to the right insertion point fast enough to change a binning decision without slowing throughput. PDF Solutions' Greg Prewitt frames the core build: collect, align, normalize, deploy the model where it is useful, and build the traceability infrastructure that makes the output trustworthy.

The practical implication: teams buying new ATE or smart-test software in the next 18 months should be evaluating the data integration story as hard as the algorithm story. A test system that cannot export device identity and process history in real time to a platform that correlates it with fab metrology is not a smart-test system, regardless of how sophisticated the ML layer is. The vendors who close that gap first (Teradyne, Advantest, or a third-party platform like Exensio) own the test data control plane. The vendors who do not are selling faster manual test.